s390x assembler pack update from HEAD.

2 files changed, 77 insertions, 36 deletions

diff --git a/crypto/sha/asm/sha1-s390x.pl b/crypto/sha/asm/sha1-s390x.pl
index 4b17848287..9193dda45e 100644
--- a/crypto/sha/asm/sha1-s390x.pl
+++ b/crypto/sha/asm/sha1-s390x.pl
@@ -21,9 +21,28 @@
 # instructions to favour dual-issue z10 pipeline. On z10 hardware is
 # "only" ~2.3x faster than software.
 
+# November 2010.
+#
+# Adapt for -m31 build. If kernel supports what's called "highgprs"
+# feature on Linux [see /proc/cpuinfo], it's possible to use 64-bit
+# instructions and achieve "64-bit" performance even in 31-bit legacy
+# application context. The feature is not specific to any particular
+# processor, as long as it's "z-CPU". Latter implies that the code
+# remains z/Architecture specific.
+
 $kimdfunc=1;	# magic function code for kimd instruction
 
-$output=shift;
+$flavour = shift;
+
+if ($flavour =~ /3[12]/) {
+	$SIZE_T=4;
+	$g="";
+} else {
+	$SIZE_T=8;
+	$g="g";
+}
+
+while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
 
 $K_00_39="%r0"; $K=$K_00_39;
@@ -42,13 +61,14 @@ $t1="%r11";
 @X=("%r12","%r13","%r14");
 $sp="%r15";
 
-$frame=160+16*4;
+$stdframe=16*$SIZE_T+4*8;
+$frame=$stdframe+16*4;
 
 sub Xupdate {
 my $i=shift;
 
 $code.=<<___ if ($i==15);
-	lg	$prefetch,160($sp)	### Xupdate(16) warm-up
+	lg	$prefetch,$stdframe($sp)	### Xupdate(16) warm-up
 	lr	$X[0],$X[2]
 ___
 return if ($i&1);	# Xupdate is vectorized and executed every 2nd cycle
@@ -58,8 +78,8 @@ $code.=<<___ if ($i<16);
 ___
 $code.=<<___ if ($i>=16);
 	xgr	$X[0],$prefetch		### Xupdate($i)
-	lg	$prefetch,`160+4*(($i+2)%16)`($sp)
-	xg	$X[0],`160+4*(($i+8)%16)`($sp)
+	lg	$prefetch,`$stdframe+4*(($i+2)%16)`($sp)
+	xg	$X[0],`$stdframe+4*(($i+8)%16)`($sp)
 	xgr	$X[0],$prefetch
 	rll	$X[0],$X[0],1
 	rllg	$X[1],$X[0],32
@@ -68,7 +88,7 @@ $code.=<<___ if ($i>=16);
 	lr	$X[2],$X[1]		# feedback
 ___
 $code.=<<___ if ($i<=70);
-	stg	$X[0],`160+4*($i%16)`($sp)
+	stg	$X[0],`$stdframe+4*($i%16)`($sp)
 ___
 unshift(@X,pop(@X));
 }
@@ -148,9 +168,9 @@ $code.=<<___ if ($kimdfunc);
 	tmhl	%r0,0x4000	# check for message-security assist
 	jz	.Lsoftware
 	lghi	%r0,0
-	la	%r1,16($sp)
+	la	%r1,`2*$SIZE_T`($sp)
 	.long	0xb93e0002	# kimd %r0,%r2
-	lg	%r0,16($sp)
+	lg	%r0,`2*$SIZE_T`($sp)
 	tmhh	%r0,`0x8000>>$kimdfunc`
 	jz	.Lsoftware
 	lghi	%r0,$kimdfunc
@@ -165,11 +185,11 @@ $code.=<<___ if ($kimdfunc);
 ___
 $code.=<<___;
 	lghi	%r1,-$frame
-	stg	$ctx,16($sp)
-	stmg	%r6,%r15,48($sp)
+	st${g}	$ctx,`2*$SIZE_T`($sp)
+	stm${g}	%r6,%r15,`6*$SIZE_T`($sp)
 	lgr	%r0,$sp
 	la	$sp,0(%r1,$sp)
-	stg	%r0,0($sp)
+	st${g}	%r0,0($sp)
 
 	larl	$t0,Ktable
 	llgf	$A,0($ctx)
@@ -199,7 +219,7 @@ ___
 for (;$i<80;$i++)	{ &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
 $code.=<<___;
 
-	lg	$ctx,`$frame+16`($sp)
+	l${g}	$ctx,`$frame+2*$SIZE_T`($sp)
 	la	$inp,64($inp)
 	al	$A,0($ctx)
 	al	$B,4($ctx)
@@ -211,13 +231,13 @@ $code.=<<___;
 	st	$C,8($ctx)
 	st	$D,12($ctx)
 	st	$E,16($ctx)
-	brct	$len,.Lloop
+	brct${g} $len,.Lloop
 
-	lmg	%r6,%r15,`$frame+48`($sp)
+	lm${g}	%r6,%r15,`$frame+6*$SIZE_T`($sp)
 	br	%r14
 .size	sha1_block_data_order,.-sha1_block_data_order
 .string	"SHA1 block transform for s390x, CRYPTOGAMS by <appro\@openssl.org>"
-.comm	OPENSSL_s390xcap_P,8,8
+.comm	OPENSSL_s390xcap_P,16,8
 ___
 
 $code =~ s/\`([^\`]*)\`/eval $1/gem;
diff --git a/crypto/sha/asm/sha512-s390x.pl b/crypto/sha/asm/sha512-s390x.pl
index e7ef2d5a9f..079a3fc78a 100644
--- a/crypto/sha/asm/sha512-s390x.pl
+++ b/crypto/sha/asm/sha512-s390x.pl
@@ -26,6 +26,26 @@
 # favour dual-issue z10 pipeline. Hardware SHA256/512 is ~4.7x faster
 # than software.
 
+# November 2010.
+#
+# Adapt for -m31 build. If kernel supports what's called "highgprs"
+# feature on Linux [see /proc/cpuinfo], it's possible to use 64-bit
+# instructions and achieve "64-bit" performance even in 31-bit legacy
+# application context. The feature is not specific to any particular
+# processor, as long as it's "z-CPU". Latter implies that the code
+# remains z/Architecture specific. On z900 SHA256 was measured to
+# perform 2.4x and SHA512 - 13x better than code generated by gcc 4.3.
+
+$flavour = shift;
+
+if ($flavour =~ /3[12]/) {
+	$SIZE_T=4;
+	$g="";
+} else {
+	$SIZE_T=8;
+	$g="g";
+}
+
 $t0="%r0";
 $t1="%r1";
 $ctx="%r2";	$t2="%r2";
@@ -44,7 +64,7 @@ $tbl="%r13";
 $T1="%r14";
 $sp="%r15";
 
-$output=shift;
+while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
 
 if ($output =~ /512/) {
@@ -78,7 +98,8 @@ if ($output =~ /512/) {
 }
 $Func="sha${label}_block_data_order";
 $Table="K${label}";
-$frame=160+16*$SZ;
+$stdframe=16*$SIZE_T+4*8;
+$frame=$stdframe+16*$SZ;
 
 sub BODY_00_15 {
 my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
@@ -93,9 +114,9 @@ $code.=<<___;
 	xgr	$t0,$t1
 	$ROT	$t1,$t1,`$Sigma1[2]-$Sigma1[1]`
 	 xgr	$t2,$g
-	$ST	$T1,`160+$SZ*($i%16)`($sp)
+	$ST	$T1,`$stdframe+$SZ*($i%16)`($sp)
 	xgr	$t0,$t1			# Sigma1(e)
-	la	$T1,0($T1,$h)		# T1+=h
+	algr	$T1,$h			# T1+=h
 	 ngr	$t2,$e
 	 lgr	$t1,$a
 	algr	$T1,$t0			# T1+=Sigma1(e)
@@ -113,7 +134,7 @@ $code.=<<___;
 	 ngr	$t2,$b
 	algr	$h,$T1			# h+=T1
 	 ogr	$t2,$t1			# Maj(a,b,c)
-	la	$d,0($d,$T1)		# d+=T1
+	algr	$d,$T1			# d+=T1
 	algr	$h,$t2			# h+=Maj(a,b,c)
 ___
 }
@@ -122,19 +143,19 @@ sub BODY_16_XX {
 my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
 
 $code.=<<___;
-	$LD	$T1,`160+$SZ*(($i+1)%16)`($sp)	### $i
-	$LD	$t1,`160+$SZ*(($i+14)%16)`($sp)
+	$LD	$T1,`$stdframe+$SZ*(($i+1)%16)`($sp)	### $i
+	$LD	$t1,`$stdframe+$SZ*(($i+14)%16)`($sp)
 	$ROT	$t0,$T1,$sigma0[0]
 	$SHR	$T1,$sigma0[2]
 	$ROT	$t2,$t0,`$sigma0[1]-$sigma0[0]`
 	xgr	$T1,$t0
 	$ROT	$t0,$t1,$sigma1[0]
-	xgr	$T1,$t2				# sigma0(X[i+1])
+	xgr	$T1,$t2					# sigma0(X[i+1])
 	$SHR	$t1,$sigma1[2]
-	$ADD	$T1,`160+$SZ*($i%16)`($sp)	# +=X[i]
+	$ADD	$T1,`$stdframe+$SZ*($i%16)`($sp)	# +=X[i]
 	xgr	$t1,$t0
 	$ROT	$t0,$t0,`$sigma1[1]-$sigma1[0]`
-	$ADD	$T1,`160+$SZ*(($i+9)%16)`($sp)	# +=X[i+9]
+	$ADD	$T1,`$stdframe+$SZ*(($i+9)%16)`($sp)	# +=X[i+9]
 	xgr	$t1,$t0				# sigma1(X[i+14])
 	algr	$T1,$t1				# +=sigma1(X[i+14])
 ___
@@ -212,6 +233,7 @@ $code.=<<___;
 .globl	$Func
 .type	$Func,\@function
 $Func:
+	sllg	$len,$len,`log(16*$SZ)/log(2)`
 ___
 $code.=<<___ if ($kimdfunc);
 	larl	%r1,OPENSSL_s390xcap_P
@@ -219,15 +241,15 @@ $code.=<<___ if ($kimdfunc);
 	tmhl	%r0,0x4000	# check for message-security assist
 	jz	.Lsoftware
 	lghi	%r0,0
-	la	%r1,16($sp)
+	la	%r1,`2*$SIZE_T`($sp)
 	.long	0xb93e0002	# kimd %r0,%r2
-	lg	%r0,16($sp)
+	lg	%r0,`2*$SIZE_T`($sp)
 	tmhh	%r0,`0x8000>>$kimdfunc`
 	jz	.Lsoftware
 	lghi	%r0,$kimdfunc
 	lgr	%r1,$ctx
 	lgr	%r2,$inp
-	sllg	%r3,$len,`log(16*$SZ)/log(2)`
+	lgr	%r3,$len
 	.long	0xb93e0002	# kimd %r0,%r2
 	brc	1,.-4		# pay attention to "partial completion"
 	br	%r14
@@ -235,13 +257,12 @@ $code.=<<___ if ($kimdfunc);
 .Lsoftware:
 ___
 $code.=<<___;
-	sllg	$len,$len,`log(16*$SZ)/log(2)`
 	lghi	%r1,-$frame
-	agr	$len,$inp
-	stmg	$ctx,%r15,16($sp)
+	la	$len,0($len,$inp)
+	stm${g}	$ctx,%r15,`2*$SIZE_T`($sp)
 	lgr	%r0,$sp
 	la	$sp,0(%r1,$sp)
-	stg	%r0,0($sp)
+	st${g}	%r0,0($sp)
 
 	larl	$tbl,$Table
 	$LD	$A,`0*$SZ`($ctx)
@@ -265,7 +286,7 @@ $code.=<<___;
 	clgr	$len,$t0
 	jne	.Lrounds_16_xx
 
-	lg	$ctx,`$frame+16`($sp)
+	l${g}	$ctx,`$frame+2*$SIZE_T`($sp)
 	la	$inp,`16*$SZ`($inp)
 	$ADD	$A,`0*$SZ`($ctx)
 	$ADD	$B,`1*$SZ`($ctx)
@@ -283,14 +304,14 @@ $code.=<<___;
 	$ST	$F,`5*$SZ`($ctx)
 	$ST	$G,`6*$SZ`($ctx)
 	$ST	$H,`7*$SZ`($ctx)
-	clg	$inp,`$frame+32`($sp)
+	cl${g}	$inp,`$frame+4*$SIZE_T`($sp)
 	jne	.Lloop
 
-	lmg	%r6,%r15,`$frame+48`($sp)	
+	lm${g}	%r6,%r15,`$frame+6*$SIZE_T`($sp)	
 	br	%r14
 .size	$Func,.-$Func
 .string	"SHA${label} block transform for s390x, CRYPTOGAMS by <appro\@openssl.org>"
-.comm	OPENSSL_s390xcap_P,8,8
+.comm	OPENSSL_s390xcap_P,16,8
 ___
 
 $code =~ s/\`([^\`]*)\`/eval $1/gem;